AkshayPanchal
AI Systems Builder
- Agents
- Retrieval
- Evaluation
- Infrastructure
I design and ship production AI: grounded retrieval, constrained agents, measurable evaluation, and the infrastructure that keeps them reliable.
Founder, Tuathra ↗, an AI engineering studio
Lead AI Engineer at Baseel (team of 12)
- Baseel
- Lead AI Engineer · team of 12
- Tuathra
- Founder
- MTU Cork
- MSc in Data Science and Analytics
- Elsevier
- Published paper
- McKinsey
- McKinsey Forward
- Experience
- 5.5 years · 15+ projects
I build the layer between a convincing demo and a system a business can depend on: grounded retrieval, constrained agents, measurable evaluation and infrastructure that stays observable at 3am.
Systems,
not screenshots.
Two systems, each documented the way an engineer would review it: the problem, the architecture, the decision that mattered, how it was evaluated, and what changed.
LexAI
A citation-first research system that reads long legal documents, reasons across them, and answers with traceable evidence.
- Problem
Keyword search returns documents; practitioners need defensible answers.
- Architecture
Hybrid retrieval over Qdrant feeding a LangGraph agent with explicit tools.
- Decision
Every claim must carry its source span, or the answer is refused.
- Evaluation
Golden question–answer–citation triples run on every change.
- Result
Answers that can be audited line by line instead of trusted blindly.
Self-healing AI SRE
A constrained multi-agent operator that detects production anomalies, reasons about root cause, and applies reversible fixes.
- Problem
Routine incidents consume on-call time and arrive as alert storms, not causes.
- Architecture
Telemetry → anomaly grouping → diagnosing agent → guarded actuator.
- Decision
Agents may only take allow-listed, reversible actions, never arbitrary commands.
- Evaluation
Every action is verified against health checks and rolled back if they fail.
- Result
Routine failures resolve and document themselves; people handle the novel ones.
Studio
Tuathra.
Alongside my role at Baseel, I run Tuathra, a small AI engineering studio that takes retrieval, agent and evaluation systems from idea to production for teams that need them to hold up.
Visit the studioFrom model
to reliable product.
Six disciplines that decide whether an AI product survives contact with real users.
- 01
Generative AI & LLMs
Probabilistic models wrapped in deterministic contracts: typed outputs, validation and fallbacks.
Model routing · structured outputs · guardrails - 02
Retrieval (RAG)
Dense and lexical retrieval fused, re-ranked, and returned with the source span it came from.
Qdrant · hybrid search · citations - 03
AI Agents
Explicit state graphs with tool budgets, stopping conditions and recoverable failure paths.
LangGraph · tool use · state machines - 04
Multi-Agent Systems
Supervisor and delegate topologies where each agent owns a narrow, auditable responsibility.
CrewAI · delegation · long-running workflows - 05
LLM Evaluation
Golden datasets and trace-level scoring that run on every change, not once before launch.
Traces · datasets · regression suites - 06
AI Infrastructure
Serving, caching, autoscaling and telemetry: the parts that decide cost and uptime.
Docker · Kubernetes · AWS · MLOps
Next step